Speech Recognition Noise Adaptation via Segmentation Voting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition devices experience decreased performance in noisy environments due to variations in noise level measurement based on the timing of speech production, leading to incorrect recognition results, especially when the signal-to-noise ratio is low.
Innovation Solution
Generating multiple pieces of speech data with varying start positions of non-speech regions by shifting the reading position from the speech region's start position by a predetermined time, and using multiple speech recognition engines to process each data set, with the most frequent recognized result being output as the final result.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the start position of the non-speech region is changed to adapt to different speech timing, then the adaptability to speech production timing is improved, but the measurement precision of noise level deteriorates because the average noise level varies depending on the start position
Solution Approach 1:
The patent divides the speech recognition process into multiple segments by creating several pieces of speech data with different start positions for the non-speech region. Each segment is processed independently through speech recognition, and the results are combined through voting to determine the final recognized result. This segmentation allows the system to adapt to different speech production timings while maintaining accurate noise level measurement for each segment.
Solution Approach 2:
The patent changes the parameter of the non-speech region start position to create multiple variations of speech data. By systematically varying this parameter (the start position), the system generates diverse speech data sets that can accommodate different speech timing scenarios. Each variation is then processed to obtain recognition results, which are aggregated to produce the final output.
2Reliability
If multiple speech recognition processes are performed with different start positions, then the reliability of recognition result is improved through voting mechanism, but the processing time increases due to multiple recognition operations
Solution Approach 1:
The speech data is segmented into multiple pieces with different non-speech region start positions, and each segment is processed through speech recognition. The recognition results from these segments are then combined using a voting mechanism to determine the final result. This segmentation approach improves reliability by distributing the recognition task across multiple processed segments.
Solution Approach 2:
The patent performs preliminary actions by pre-processing the speech data into multiple pieces with varied start positions before the actual speech recognition process. This preliminary segmentation and preparation of multiple data variants enables the subsequent recognition processes to work more efficiently and the voting mechanism to operate on pre-prepared results, reducing overall processing time.
3Measurement precision
If the non-speech region length is extended to ensure sufficient noise measurement, then the measurement precision of noise level is improved, but the loss of speech information increases due to deletion of speech region
Solution Approach 1:
Instead of using a single long non-speech region for noise measurement, the patent segments the speech data into multiple pieces, each with its own non-speech region. This allows for sufficient noise measurement in each segment without requiring an excessively long non-speech region that would delete significant speech information. The segmentation distributes the noise measurement task across multiple smaller regions.
Solution Approach 2:
The patent applies partial action by using multiple smaller non-speech regions distributed across different speech data pieces, rather than one excessive long non-speech region. Each non-speech region provides sufficient noise measurement for its corresponding speech segment, and the combination of these partial measurements achieves the overall goal without excessive deletion of speech information.
Data Source
AI summary
Speech recognition performance is improved without changing a speech recognition engine. A speech data generation section generates, from speech data for which speech recognition is to be performed, a plurality of pieces of speech data whose starting positions of the non-speech regions differ. A speech recognition engine performs speech recognition by using each of the pieces of speech data. A totaling/comparison section provides the most numerous recognized result from among a plurality of obtained recognized results.